Lasse Cronqvist - Using Multi-Value Logic in Social Science

Using Multi-Value Logic Synthesis in Social Science

Paper prepared for the 2nd General Conference of the European Consortium for Political
Research (ECPR), Marburg, September 2003. Section 6: Methodological Advances in
Comparative Research : Concepts, Techniques, Applications, Panel 8: Assessing the
respective potential of Qualitative Comparative Analysis (QCA), Fuzzy Sets and other
techniques : applications.
Lasse Cronqvist
Institute of Political Science
Philipps-University,
Marburg/Germany
E-Mail: lasse@staff.uni-marburg.de

Abstract:
One of the main problems using Qualitative Comparative Analysis has been the restriction to
use only dichotomised variables. In this paper an extension to QCA is presented, which allows
the use of multi-valued variables with Boolean synthesis, and the problem of setting the
thresholds is discussed to overcome this methodological problem. A new software named





(Tool for Small-N Analysis) is introduced which allows the use of Multi-Value

QCA (MVQCA). Also some of the problems of setting threshold used with QCA and MVQCA
are discussed.

Lasse Cronqvist - Using Multi-Value Logic in Social Science

1.

Introduction1

QCA has been used for research on several topics of political science and abroad (see DeMeur
/ Rihoux: (2002) for a listing of some examples), but the method also has its shortcomings.
One of the main problems with QCA has been stated as the compulsory transformation of data
to dichotomy data (Berg-Schlosser 2002), which bears the risk of data loss and may create a
large number of contradicting case configurations, which diminish the use of the analysis.
In FS/QCA Ragin and Drass (2002) implemented the fuzzy-set approach presented by Ragin
(2000) to overcome the dichotomy problem. With fuzzy-sets it is possible to assign values
between 0 (fully out) and 1 (fully in) to each variable to a case, giving the possibility to create
a more detailed case configuration. These configurations are then processed with proportional
methods, to find the necessary and sufficient conditions to explain an outcome.
In this paper, an other approach for the use of non-dichotomy data in social science is
introduced: The Multi-Value Qualitative Comparative Analysis (MVQCA) keeps the idea of
minimizing expressions with Boolean Analysis by allowing non-Boolean literals to be included
in the analysis. This means that a truth function is assigned for each literal, so that the literals
can be combined like Boolean literals in QCA are being combined. This allows the use of
Multi-value scales within one variable, and then to use them to minimize complex data sets to
parsimonious expressions. MVQCA thereby allows the use of Multi-value variables as they are
also used in many fuzzy-set applications (for example Ragin 2000: 271-307, Pennings 2003),
but instead of apply an analysis using probabilistic criteria to identify sufficient and necessary
conditions, the veristic synthesis of data sets used in QCA is still used with MVQCA.
MVQCA uses the ideas of Multi-value Logic Synthesis developed by the electronic engineers
at Berkeley University (see Brayton / Sunil 1999), which have used Multi-value synthesis for
computer aided circuit design. Their papers on this topic are a good help to access the
mathematical basics of the topics. Although there is also a software program provided by the
researchers at the Berkeley University (see MVSIS Group), a new software called


 

(Tool for Small-N Analysis) has been developed for the use of MVQCA (Cronqvist 2003a).

Acknowledgement: I wish to thank Prof. Dr. Dirk Berg-Schlosser for his many helpful comments on the draft
of this paper.

Lasse Cronqvist - Using Multi-Value Logic in Social Science

This is mainly due to guarantee that further developments on (MV)QCA as well as other
methods of macro-qualitative social science can be implemented easily, and because specific
topics related to social science data analysis do not exist in electronic engineering applications
(like ordinal variables just to mention one).

This paper
In his paper on the shortages of the case-oriented approach as well as of the variable-oriented
approach, Goldthorpe (1997) explains that the limitation of QCA to dichotomous coding of
data is a further restriction to social research. The high sensitivity of calculation toward the
threshold setting inflicts the power of the analysis in a severe way (Goldthorpe 1997:3), and, to
use the QCA-language, a slightly change of the threshold may result in a different solution
of the analysis. In fact, using unlucky thresholds can place the scientist on a dead end road, if
the result of the data analysis is used without further analysis.
In this paper I want to introduce an extension of QCA with multi value variables. I will start
with a short introduction of the subject of the subject of Multi-Value Qualitative Comparative
Analysis (MVQCA). Without going into deep technical details, the basic features of this
method will be explained, and these will be illustrated by some examples. Secondly, the
potential of MVQCA to diminish the consequences of re-coding interval scaled variables to
ordinal scales with only a few scale values, will be shown by adapting a data set earlier used
for QCA analysis to MVQCA methods, which can reduce the number of contradictions
considerably. Finally, I will make some remarks on "threshold setting" - the question how
thresholds can be set in a responsibly way. All this thoughts are also present in the new
software for (MV)QCA -


 
.




software can be found in the

More information on how to use the

User Manual (Cronqvist 2003b).


 

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Multi Value Qualitative Comparative Analysis (MVQCA)

First, the main features of the MVQCA method will be described. Readers familiar with the
Qualitative Comparative Analysis developed for the social sciences by Charles Ragin (see
Ragin 1987) will discover some similarities, but also some fundamental differences between
these methods.
Basic Features of MVQCA
Multi-Value analysis requires that each variable is available on either an ordinal or a nominal
basis. As an example, the variable region , which contains information about the geographical
location of a country, is a nominally scaled variable, while the variable GDP-Growth with
the possible values of PGDP={negative, low or zero, medium, high, very high} is an example of
a five-step ordinal scale. This means that raw interval scaled data can not be used for MVQCA
directly, but must be transformed into values inside a multi value set. This is done by selecting
an array of thresholds according to which the values are assigned. This, of course, implies a
certain loss of more finer-graded information. This means that there may be no differences
between two cases in the transformed data, although there might be a (more or less small)
difference in the original raw data. Therefore the choice of thresholds has to be made very
carefully, a subject to be discussed in depth further below.
Multi-Value Notation
As multi-value variables can have more than two states, set notation has to be used to represent
the cases and implicants as formal expressions. Each expressions consists of one or more
literals XiSn where Xi is a variable from the data set and Sn is a set of values of the variable valid
in this expression. I=A{0,1} indicates that the implicant I represents all cases having a value of
A which is either 0 or 1. To make the reading more comfortable, the notation I=A0,1 is used in
this text as well. The notation I=A0-2 may be used to indicate A having all values between 0 and
2 (= 0,1,2), and I=A~2 means all values except for 2.2

Researchers used to Boolean QCA will see that this notation can also be used to express Boolean terms: Instead
of using upper case letters (A) for indicating the presence of a condition and lower case letters (a) for the absence
of a condition, the value 0 (A0) will be used for indicating the absence of a condition and the value 1 (A1) for the
presence of a condition.

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Combining literals and implicants.

Very rarely a data set can be explained by just one literal. Almost always explanations have to
be found in combined implicants, consisting of more than one literal, or even multiple,
combined implicants. For this Boolean algebra is used. Each literal XiSn represents a number of
cases, which have one of the states in the variable Xi indicated by Sn. For those cases Ci the
Boolean Truth function is XiSn(Ci) = T (true), whilst is F (false) for other cases. The
following simple example will illustrate this:
Case Number
1
2
3
4

Value of A Value of B Value of C

0
1
1
1
2
1
2
2
1
0
1
2

A{0}(C)
T
F
F
T

Table 1: Example 1. Multi Value Cases and the Boolean Truth function

The Boolean Truth function of the literal A0 is True for the cases 1 and 4 as the value of A
those cases is exactly 0, while the function is False in the cases 2 and 3 as the value of A is not
0.
Boolean AND
To combine multiple literals into one expression, the Boolean AND is used. This is sometimes
called Boolean multiplication (Ragin 1987: 91), which may be confusing as it is substantially
different from arithmetic multiplication. If an expression consists of two literals combined by
the Boolean and (the sign is used here), the expression is only true if the Boolean Truth
function of both literals with the specific case is true. This means that in the upper example the
Boolean Truth function of the expression A0C1 is only true for case 1, as only case 1 is
having the value 0 for variable A and 1 for the variable C. In standard use of MVQCA, the sign
is mostly left out, and the expression A0C1 is written as A0C1.
AB

AB

Figure 2. Boolean OR

Figure 1. Boolean AND

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Boolean OR
To combine two or more Boolean expressions, the Boolean OR is used, which sometimes also
is called Boolean addition. If a Boolean expression consists of two Boolean expressions
combined by OR (), the Boolean Truth function of the combined expression is true if one or
both Boolean Truth functions of the expressions used to combine the new expression is true. In
example 1 the Boolean Truth function of the combined expression A0B1 A1B2 is True for the
cases 1,2, and 4 as the first expression of the addition is true for case 1 and 4 and the latter for
case 2.
In daily use, the OR sign can be replaced by the normal addition + sign, if it is clear that the
Boolean addition is used.
Representing Data
To use Multi-Value coded data it is necessary to construct the data as a data matrix. Unlike the
truth tables used by QCA, the data variables used are not restricted to binary, nominal-scale
variables (Ragin 1987: 87), but can be non-binary and of an ordinal type, too. Also, not all
possible logical combinations are represented in the matrix. In fact this would very quickly
exhaust the available resources, as we would run into a very high number of possible
combinations. Considering a simple example should make this clear:
As an example for Qualitative Comparative Analysis Lipsets social indicators explaining the
democratic breakdown in Interwar Europe have been used (Berg-Schlosser and De Meur 1994:
255-257). Lipset employs four socio-economic indicators sustaining democracy. These are:
Gross National Product (G), Degree of Urbanisation (U), Literacy (L), and Industrial Labour
Force (I). If the variables are transformed into Boolean variables, there are 16 (2*2*2*2)
logical combinations. In fact, the number of logical combinations in a truth table is given by
the number of variables, and calculated with the formula =2|v| where |v| indicates the number
of variables. In MVQCA the number of combinations is higher for the same number of
variables, as every variable may have more than two values assigned.3

Please note that the range of values can be selected for each variable independently, and that they may differ
from each other. For MVQCA, the numbers {0,..,n-1} have to be used to construct the range of the size n.

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Variable name
GDP
Urbanisation
Literacy
Industrialisation

Possible values (P)

Number of Variables (vi )
very low, low, medium, high, very high
5
very low, low, medium, high, very high
5
low, medium, high, almost 100%
4
low, medium, high
3

Table 2: Number of Variables in Multi-Value Sets

In table 1.1. the Lipset indicators are assigned (exemplary) Multi-Value ranges. In this
example there are 300 (5*5*4*3) possible combinations to deal with. The number of
combinations is given by the formula =|1| x ... x |n|, meaning that you have to multiply the
number of possible values of each variable to get the number of possible combinations. In fact
it is possible to recode multi value sets to binary sets using dummy variables4 but given the
very high number of possibilities resulting from this recoding, this approach seems to be less
useable.5 Also, the limitation to a rather small number of variables in the common software
used for QCA does not allow the usage of middle size MVQCA data sets.6 Therefore a new
algorithm has been included in the


 

Software Package (Cronqvist 2003a), which is

using a multi-value set algorithm to calculate the minimized expressions (see Cronqvist 2002).
Parsimony
The aim of MVQCA is (similar to QCA) to find parsimonious explanations (Ragin 1987: 83)
in complex data sets, and thereby to specify variables of major importance. Unlike

This possibility is given by the Espresso algorithm included in FS/QCA (Ragin/Drass 2002).
In fact, there are a number of problems by encoding multi value data to binary data (Brayton and Khatri 1999:
3): The main disadvantage is the number of additional variables added, by which the number of dont cares rises
to very high levels. This is mainly because a lot of combinations are created which will never occur. Imagine a
traffic light: To make this example easy, the traffic light has three possible values: Red light, yellow light, and
green light. Encoding the Multi-Value variable Light with PLight={red,yellow,green} into three binary dummy
variables (red,yellow,green) would result in the following table:
5

MV Value
Red
Yellow
Green

red value
1
0
0

yellow value
0
1
0

green value
0
0
1

Now the point is that not only the upper (real) combinations will be created, but also all other possibilities will be
constructed. In reality, a combination like red=1 and green=1 will never occur in regular traffic light usage, but
the combination still occurs in the data matrix when including remainder cases in the analysis. This means that a
high number of never used combinations will be created and combined with the other independent variables, and
these must all be included in the analysis.
6
In QCA3.0 (Drass/Ragin 1992) the number of variables was restricted to 12. In FS/QCA (Ragin/Drass) there is
no restriction to the number of variables included, but the software produces an error if more than 25 variables
are used. Of course this is no harm when dichotomious variables are used, as this number of variables would not
be used often. But when using dummy variables for multi-value reduction, the restriction to 25 dummy variables
would mean a restriction to a smaller number of variables, as every multi value variable has to be recoded to
multiple binary variables.

Lasse Cronqvist - Using Multi-Value Logic in Social Science

quantitative, statistical methods, it is possible to discover new connections between variables

instead of only checking old connections. (Ragin 1987: 84). The main principles of MVQCA
will be described by two imaginary data sets:
Variable name
A
B
C
Outcome

Possible values (P)

0,1
0,1,2
0,1,2
0,1,2

Number of Variables
2
3
3
3

Table 3. Variables in example data set (Example 2+3).

Case
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1

0
0
0
1
1
1
2
2
2
0
0
0
1
1
1
2
2
2

0
1
2
0
1
2
0
1
2
0
1
2
0
1
2
0
1
2

Outcome
0
0
0
0
1
1
2
1
2
1
0
1
1
1
2
1
2
2

Table 4. Example data set (Example 2) .

Minimization
The goal of MVQCA is to extract short (parsimonious) explanations of an outcome by
minimizing the complex multi-value data set. In QCA, minimization is rather easy: If two
expressions only differ in one term, these two can be taken to produce a combined, reduced
expression. If two logical combinations both have the outcome 0 and their expressions are
given by abc and Abc, then A is irrelevant, and the combined expression can be stated as bc. As
indicated above, multi-value synthesis is a generalisation of Boolean synthesis: A number of
expressions can only be replaced by a reduced expression if all expressions in the data
including the reduced expression have the same outcome.7 In example 2 the expression
A0B0C0, A0B0C1, and A0B0C2 can be reduced to A0B0 because all three expressions in the data
sets implied by this reduced expression have the same outcome 0.
7

If an expression is included in , this means that all variable values included in are also included in .
Then it can also be said that implies .

Lasse Cronqvist - Using Multi-Value Logic in Social Science

In other words, the most fundamental rule of Boolean reduction as expressed by Ragin can be
rewritten for multi-value reduction:
(1) If two Boolean expressions differ in only one causal condition yet produce the same outcome, then
the causal condition that distinguishes the two expressions can be considered irrelevant and can be
removed to create a simpler, combined expression. (Ragin 1987: 93).

For multi-value reduction this can be written as: 8

(2) If all n multi-value expressions (c0,...,cn-1) differ only in the causal condition C while all n
possible values of c yet produce the same outcome, then the causal condition C that distinguishes these
n expressions can be considered irrelevant and can be removed to create a simpler, combined
expression .

The reader should convince himself that the rule for Boolean reduction is a specialisation of
the rule for multi-value reduction, and that the rule for multi-value reduction also is valid for
Boolean reduction.
But as the number of don t cares will be very high in data sets with many fine-graded scales,
minimizing cases with the same outcome without including don t cares would be a very short
event, if such a data set is used. To remove a causal condition requires that all expressions
implied by this condition produce the same outcome. In the Lipset data set mentioned above,
this would mean that to remove G, all expressions with G implied by the new specified
expression UxLyIz would have to have the same output. This might be possible for one single
combined expression, but as minimizing should move beyond the point of single reductions
and we want to reduce already combined expressions, this strategy seems non practicable: You
would need at least 25 (5 possible values for G * 5 possible values for U) cases to be able to
get a combined expression of LyIz with the same outcome. Considering the low number of
cases in small-N Analysis this shows the problem clearly. Therefore, cases that exist logically
but not in the data set - the remainder cases will be included in the synthesis. This creates a
new rule for multi-value reduction including remainder cases:
(3) If all n multi-value expressions (c0,...,cn-1) differ only in the causal condition C with n possible
values yet produce the same outcome or are non existing expression (remainders), then the causal
condition C that distinguishes these n expressions can be considered irrelevant and can be removed to
create a simpler, combined expression .

indicates an combined expression.

Lasse Cronqvist - Using Multi-Value Logic in Social Science

But as there are only three groups of expressions (Cases with outcome o, Cases with an
outcome other than o, remainders) 9, this means the rule can be changed to:
(4) If two or more multi-value expressions ci   { c0 ,..., c|n-1| } differ only in the causal condition C
with n possible values yet produce the same outcome, then the causal condition C that distinguishes
these n expressions can be considered irrelevant and can be removed to create a simpler, combined
expression, if there is no expression implied by the new expression producing a different outcome.

In fact, even if only one expression is implied by the expression multiplied with C, C can be
removed if no expression implied by is producing a different outcome.
Case
1
2
3
4
5
6
7
8
9

A
0
0
0
0
0
1
1
1
1

B
0
0
0
1
2
0
0
2
2

C
0
1
2
0
2
0
1
0
1

Outcome
0
0
0
0
2
1
0
1
2

Table 5 Example data set (Example 3)

In Example 3, there is a number of combinations with no outcome assigned (don t cares),

which are not stated. For the outcome 0 using the upper rule (4), we find that the cases 1-3 can
be reduced to A0B0. Also, case 7 (A1B0C1) can be reduced to B0C1 as there is no case with an
other outcome implied by this expression. For the outcome 0, the following reductions are
possible:
A0B0C0, A0B0C1, A0B0C2

can be reduced to A0B0

A0B0C0, A0B1C0

can be reduced to A0C0

A0B0C1, A1B0C1

can be reduced to B0C1

A0B0C1

can be reduced to A0C1

A0B0C2

can be reduced to A0C2

A0B1C0

can be reduced to A0B1

A0B1C0

can be reduced to B1C0

It is even possible to reduce A0B1 and B1C0 to B1.

9

Readers with knowledge of QCA may object that there are also combinations with contradictory (different)
outcomes. This is, of course, also true in MVQCA, but here such outcomes are given a special outcome value
and are viewed as an independent group of cases, which may be merged with any other group or ignored, if
desired.

10

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Excluding Remainders
When working with data sets only containing few non dichotomous variables, it also makes
sense to exclude remainders from the reduction sometimes. Then an adopted Quine algorithm,
which is following the rule set above as (2) is performed over the data set. In Example 3 the
cases 1,2, and 3 can be merged, as the three cases only differ in the Variable C and they all
have the same outcome. Also, all possible cases implied by the new expression resulting from
the reduction are existing (and having the outcome 0). Hence it can be noted that
A0B0C0

+ A0B0C1 + A0B0C2

A0B0

As stated, this method will show useable results with data sets with only a few multi value
variables only. If too many multi valued variables are used, there will be no or only a very
limited number of reduction steps possible.

Implicants, Prime Implicant, Cover and Prime Cover

The idea of implicants and covers has already been used by Charles Ragin for the construction
of QCA (Ragin 1987). The same concept is used with MVQCA. The goal is to find a minimal
prime cover which covers all the cases with an outcome o without covering any case with an
other outcome than o. To understand this, first, some definitions have to be made.10
Map: A case C is mapped by an expression  , if  is a subset of the expression of C (  implies
the expression of C).
Set of Representation: The set of representation of an expression

consists of all cases

mapped by  .
Implicant: An implicant is a multi-value expression, with all cases from the data set mapped in
it have the outcome chosen. [= no case mapped in the implicant may have another outcome]
Prime Implicant: An expression

is a prime implicant if there is no other implicant  which

implies  .
Cover: A cover is a set of implicants whose union of case memberships contains all the cases
with the selected outcome.

10

A mathematical definition of these definitions is given by Brayton and Khatri (1999). Readers with some indepth knowledge of (formal) mathematics should take a look at the formal definition of multi value logic
synthesis.

11

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Prime Cover: A prime cover is a cover whose elements are all prime.
Minimal Prime Cover: A prime cover is minimal if no prime implicant can be removed from
the prime cover without breaking the cover quality of the set of prime implicants.
The goal of minimizing complex multi-value data sets is achieved when a prime cover is
found. To find such a prime cover, two steps of calculations are necessary: The first step has to
find the prime implicants, whilst the second step has to combine these prime implicants to
covers as small as possible.11 As the method used in QCA, which reduces the data set from top
to bottom, would imply a huge amount of memory to be used to process the calculation when
using multi-value sets, another algorithmic approach has been chosen to find the prime
implicants. For example, consider again the Lipset indicators (see table 2). This data set has a
complexity of 5*5*4*3 = 300 possible configurations in the data matrix. With modern
computer technology processing this may not be a problem. But only few data matrices consist
of only four independent variables, often ten or more variables have to be taken into account
when using MVQCA. A data set with ten variables with two variables with four values, five
variable with three values and three dichtomized variables gives a mass of 42*35*23 = 31104
possible configurations. Then nearly all possible configurations have to be reduced, if the
remainders are included in the analysis. So if we have a data set consisting of four variables
like the Lipset indicators, and we are looking at eight cases with the outcome 0 and ten cases
with the outcome 1, we have to reduce 292 cases having the outcome 1 or being don t cares to
find the prime implicants for the outcome 1. If we are having the same number of cases for a
ten variables data set, we are looking on 31096 cases to be reduced.12 To reduce multi-value
expressions, therefore, a new algorithm has to be used, which finds the prime implicants from
bottom to top. Briefly stated, the calculation is done starting with the shortest possible
expressions (consisting of just one literal) and extending them until they only map cases with
the selected outcome and no other cases13.
Before the solutions are found next, some calculations are done to reduce the number of prime
implicants. Let A and B be two nominal scaled variables. If A0B0 and A2B0 are both prime
11

This is the same procedure as in Boolean QCA (Ragin 1987: 95ff)

Also, it is not possible to use De Morgan s law to reduce multi-value expressions, as this law only is
applicable to Boolean data sets.
13
A Powerpoint presentation on this algorithm can be found on the author'
s web side (Cronqvist 2002).
12

12

Lasse Cronqvist - Using Multi-Value Logic in Social Science

implicants, then these are joined to the new prime implicant A0,2B0. This could also be done
with ordinal variables, but it makes little sense to combine the two prime implicants if A1B0
implies cases with an other outcome, so in





values of ordinal variables are only

joined, if they can be connected by -. In this case A0B0 and A2B0 will only be joined if
A0-2B0 is a prime implicant. This means that A1B0 must be a prime implicant its self or that
there is no case implied by A1B0. This is the point mentioned in the introduction where ordinal
and nominal variables are treated differently.
When the prime implicants have been found and eventually reduced, these have to be
combined to the shortest possible covers.


 

is able to find all shortest prime covers

of the smallest size s and can even be configured to find all prime covers, or just prime covers
up to a certain size. In MVQCA (as well as in QCA) the size is calculated by adding the size of
the prime implicants, which are calculated by the number of literals included. This means that
the size of the implicant A0B0C1 is the same as the size of the implicant A0B0C1,2 although the
latter implicant includes two values of the variable C. This way of calculating the size of a
prime cover has to be discussed, but the software (





) is constructed to adapt other

modes of measuring the size of prime implicants and prime covers.

13

Lasse Cronqvist - Using Multi-Value Logic in Social Science

2.

Improving QCA Analysis with Multi-Valued Scales

To indicate the use of MVQCA, I would like to show an example where changing the variable
scale gives a more satisfactory result.
Vanhanens Indices on Democratisation
One of the first applications of QCA in social sciences was published by Dirk Berg-Schlosser
and Gisle Demeur (1994). The authors revisited major hypotheses on the conditions of
democracy in interwar europe and made them subject to a Boolean analysis. One of the
hypothesis tested were the Indices of Democratisation by Vanhanen.
In his work on the emergence of democracy in 119 states, Vanhanen (1984) defines three
variables, which "indicates three different dimensions of politically relevant power resources"
which "together [...] measure the social conditions and structures on which the nature of
political systems depend" (Vanhanen 1984: 36). The three variables are:
Index of Occupational Diversification (IOD): This index contains the percentage of urban
population and the percentage of non-agricultural population. This variable is "assumed to
indicate the distribution of economic and occupational interests and, indirectly, the distribution
of economic and organizational power resources" (Vanhanen 1984: 36)
Index of Knowledge Distribution (IKD): This index contains percentage values of the number
of students and literates. This index should "indicate relative differences between the countries
in the distribution of intellectual power resources" (Vanhanen 1984: 36).
Share of Family Farms (FF): This value indicates the share of family farms of the total area of
holdings. (Vanhanen 1984: 24).
Berg-Schlosser and De Meur (1994: 257f) examine the explanatory potential of these three
indicators to explain the differences in the survival of democracy in 17 states of Interwar
Europe14. Table 6 shows the original data given by Vanhanen and the dichotomised data by
Berg-Schlosser / De Meur. For each variable a threshold of 50% is used, as these levels where
suggested by Vanhanen (Berg-Schlosser / DeMeur 1994: 258).
14

These cases are (with the abbreviations used further on): Austria (AU), Belgium (BE), Czechoslovakia (CS),
Finland (FI), France (FR), Germany (GE), Greece (GR), Hungary (HU), Italy (IT), the Netherlands (NL), Poland
(PL), Portugal (PO), Romania (RO), Spain (SP), Sweden (SW), Great Britain (GB). Althought Vanhanen
provides data on ten more European states the authors do not use them, as they argue that the major cases are
included in the analysis by using the selected cases (Berg-Schlosser / De Meur 1994: 254).

14

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Case
AU
BE
CS
FI
FR
GE
GR
HU
IT
NL
PL
PO
RO
SP
SW
GB

IOD
51,5
64
38,5
21,5
48
53
34
37
38
61
17,5
30,5
16,5
35
39,5
78,5

D(IOD)
1
1
0
0
0
1
0
0
0
1
0
0
0
0
0
1

IKD
55
51,5
49
46,5
50,5
54
28
47
39,5
51,5
37,5
18,5
25
33
52,5
50

D (IKD)
1
1
0
0
1
1
0
0
0
1
0
0
0
0
1
1

FF
45
30
40
47
35
54
28
40
22
40
53
20
41
20
50
25

D (FF)
0
0
0
0
0
1
0*
0
0
0
1
0
0
0
1
0

Outcome
0
1
1
1
1
0
0
0
0
1
0
0
0
0
1
1

Table 6: The data represents the values of the decade 1920-1929 for each variable in Vanhanen 1984. The
dichotomised data is from Berg-Schlosser / De Meur 1984. * Although Greece has a value of 28% of
Family Farms, the dichotomised value of 1 has been used for the Boolean analysis in the analysis by BergSchlosser and DeMeur. This error does actually not effect the results of the analysis, as Greece would join
the group of contradictonary cases with all values = 0 and Poland would still be form a own group so that
the QCA settings would not change. In this paper I will use the correct data.

The decision to use a 50% threshold for all three variables was an unfortunate one, as we will
see. When performing the Boolean synthesis, first we find that 12 of the 16 cases are
contradicting so that only two cases with the outcome 0 and two cases with the outcome 1 are
explained.

Cases
RO, FI, PO, GR, SP,
HU, IT, CS
PL
SW
FR
AU, BE, NL, GB
GE

IOD
0

IKD
0

FF
0

Outcome (QCA)
C

0
0
0
1
1

0
1
1
1
1

1
1
0
0
1

0
1
1
C
0

Table 7: Truthtable of the Vanhanen Indices using dichotomous data.

This is indeed not satisfying, and sets up a big question sign if these indices can be used to
explain democracy breakdown in the interwar period in Europe. But if we look at the
distribution of cases and the thresholds, we see several irregularities. As an example the
variable IOD:

15

Lasse Cronqvist - Using Multi-Value Logic in Social Science

45

Figure 3: Distribution of values of IOD. The thick line indicates the median, the narrow lines the new
selected threshold of 45% and the previous threshold of 50%.

Setting the threshold at 50% looks very randomly, as it separates two values close to each other
(48%, 51,5%) and it also cuts the set of cases into the subsets of different size. A simple
average linkage cluster moves the threshold to somewhere between the cases with the values of
39,5% and 48% (setting the threshold to 45% satisfies this).15
Also the threshold of 50% for IKD seems unfortunate, as it cuts right through a group of cases
close to each other:
45

Figure 4: Distribution of values of IKD. The thick line indicates the median, the narrow line the original
selected threshold of 50% and the new threshold at 45%.

Setting the threshold to 45% makes the division look more reasonable.
Finally the threshold of FF is difficult to place. Using a threshold of 50% cuts the set into two
unbalanced subsets (containing 4 and 12 cases), but due to the distribution all along the range
between 20% and 55%, it looks difficult to place the threshold without a more serious loss of
15

Setting the thresholds using a cluster may also seem at least questionable, as no theoretical assumptions about
the variable are included in the decision. But thresholds have to be set so that they clarify as much as possible
and not create artificial groups of cases. See the chapter at the end on thresholds for more on this.

16

Lasse Cronqvist - Using Multi-Value Logic in Social Science

data precision. To solve this problem, a set of two thresholds are used. They are placed at 32 %
and 43 % (using simple average clustering), which cuts the data set into three equal size
portions: One subset with low values of FF (<32%, 5 cases), one subset with middle values
(>32% and <43%, 5 cases) and one subset with high values (>43%, 6 cases).
50

Figure 5: Distribution of values of FF. The thick line indicates the median, the narrow line the two
thresholds at 32% and 43%. The dotted line indicates the original used threshold of 50% as used by BergSchlosser and De Meur.

Using these thresholds, only two cases contradict (CS and HU), whilst the other 14 cases are
not contradicting with cases of an other outcome.
Cases
PO, GR, SP, IT
RO
PL
HU, CS
FI, SW
BE, GB
FR, NL
AU, GE

IOD
0
0
0
0
0
1
1
1

IKD
0
0
0
1
1
1
1
1

FF
0
1
2
1
2
0
1
2

Outcome (QCA)
0
0
0
C
1
1
1
0

Table 8: Condensed data table with the use of a multi value coding of FF and revised thresholds of IOD
and IKD.

If we take a closer look on the raw data of the two cases contradicting, we see that they indeed
are very similar, so that dividing the cases would only be possible with a very artificial
threshold16:

Case
HU
CS

IOD
37
38.5

IKD
47
49

FF
40
40

Outcome
0
1

Table 9: Raw data for the two cases HU and CS still contracting with the new thresholds.

16

We would have to define a threshold which cuts exactly between 37% and 38,5% on the IOD scale, not a very
reasonable setting with the data used here (see chapter on how to set thresholds below).

17

Lasse Cronqvist - Using Multi-Value Logic in Social Science

By changing the thresholds we have reduced the number of cases in the contradictions from 12
to 2. This in fact raises the explanatory power by the reduced expressions, as only one case is
not included in the solutions found for the cases with the outcome 0 as well as for the cases
with the outcome 1. If we reduce the new data set, we find that the cases with the outcome 0
including the remainder cases are reduced to IKD{0}+IOD{1}FF{2}, meaning that all the cases
with a low knowledge distribution (below 45%) broke down in the interwar Europe. This is the
majority of breakdown cases (PL,GR,IT,PO,SP,HU,RO), The remaining two cases (AU,GE)
are described by a high level of occupational distribution (>45%) and a high percentage of
family farms.
On the other hand, the group of cases O(1) (Cases with the outcome 1) is reduced to
IOD{1}FF{0,1}+IOD{0}IKD{1}FF{2}. The most cases (FR, NL, BE, UK) had a high
occupational distribution combined with a low or medium percentage of family farms. The two
cases FI and SW could not be reduced further, but their combination of a low occupational
distribution, a high knowledge distribution and high percentage of family farms identifies them
clearly as cases with the outcome 1.

3.

Finding "good" thresholds

In this paper the idea of the extension of QCA by Multi-Value Scales has been drafted. It is
important to realise, that MVQCA unlike QCA uses non Boolean Data Representation, but that
the two step approach of MVQCA is the same as the one of QCA: In the first step, the
necessary prime implicants are calculated, and these are then combined to sufficient and
necessary solutions in the second step. As the choice of thresholds is substantial for the success
of the analysis some aspects on selecting thresholds will now be touched, which have to be
discussed in the further development of the QCA method:

One critique on QCA has been that the choice of thresholds can decide the result and
therefore scientist can be tempted to adjust their thresholds so they gain the best results
possible. This raises the question on the legitimacy of threshold manipulation as it is used
above. Thresholds are used to separate data points representing phenomenons in the real
world . These thresholds should be chosen so that they create natural subsets of all data
points. This means that all data points have to be fitted into as homogenous subsets as
18

Lasse Cronqvist - Using Multi-Value Logic in Social Science

possible and that artificial cuts (e.g. between two near-by points) should be avoided.





provides a Thresholdsetter a visual aid for this task, where the thresholds

can be set either by a calculated cluster subsection or be manipulated by hand. It is

important that all thresholds should be set only looking at the data of the processed
variable, and that the values of the outcome variable should have no influence at all on the
threshold of the independent variables.

It may appear strange to use a threshold of 45% instead of 50%. But it must be kept in
mind that these values are all only numbers expressing a real world phenomenon. Due to
the subject dealt with in social science, these are affected by a certain degree of
uncertainty. This implies that 50% must not necessarily be a more appropriate threshold
than 45%. Threshold are used to separate data from other data. Therefore thresholds should
be chosen so that they are discriminating as much as possible (without being artificial),
creating as homogenous data sets as possible. Therefore thresholds will often not to be
convenient numbers (like 50%, 10.000  etc.).

The used thresholds also have to be mentioned when using (MV)QCA results: Stating that
cases are having a low value with the Variable FF does not mean the same with different
thresholds. A threshold of 45% may produce a different set of subsets as a threshold of
50% (A case with 48% is below the threshold of 50% but above the threshold of 45%).
Therefore, cases with a low value should correctly be addressed as cases with a value
below the threshold of 45% (or 50%) indicating the used threshold.

It is possible to use as many thresholds as needed to remove all contradictions. But using
too many thresholds will singulize cases so that all cases are represented by one
combination at their own. Such data sets will not provide useful solutions, as it will be
difficult to find parsimonious solutions within a set of singulized cases. The ideal threshold
definition will use as grossly scaled variables as possible and explain as many cases as
possible without contradictions. A dichotomous data set without contradiction with only a
few number of variables would be the optimal data set for the further application of QCA
results. The use of strictly dichotomious data sets will only be possible in few researches.
But as the Vanhanen example given above shows, that it is possible to exclude all
19

Lasse Cronqvist - Using Multi-Value Logic in Social Science

contradictions possible to exclude17, by changing two thresholds and adding one more
threshold to a variable. Therefore major advances seems possible with just a few wisely
chosen steps.

4.

Conclusion

Extending QCA with multi valued scales offers the opportunity to reduce the number of
contradictions in data sets used for Boolean synthesis. But the example given above also shows
the importance of the use of well chosen thresholds. In fact, only a good combination of a
good number of coding possibilities and wisely selected thresholds can bring up a recoded data
set, which can be applied successfully for Small-N comparative research.
One critical point of QCA applications is the choice of thresholds. In many applications of
QCA the thresholds are chosen from a theoretical foundation, but the consequences of the
choice are not checked, which may make the choice non defensible. A threshold set to a round
number may appear reasonable, but if it cuts just through a number of closely related cases an
artificial division is made, which is the source for questionable results.


 

provides a

couple of tools based on the ideas presented in this paper to find more suitable thresholds. But
they can only be a first step in further discussions on how to set thresholds for the use with
(MV)QCA. As mentioned in my introduction, faulty thresholds have been a major point of
critique towards the QCA approach. Using more functional thresholds improves the use of
(MV)QCA in the research process. Therefore the subject on choosing "good" thresholds should
get more attention in further development of macro-qualitative comparative analysis tools.

17

The two last cases could not be separated as their original data was to close to separate in a non-artificial way.

20

Lasse Cronqvist - Using Multi-Value Logic in Social Science

5.

Bibliography

Berg-Schlosser, Dirk and Gisle DeMeur (1994): Conditions of Democracy in Interwar

Europe: A Boolean Test of Major Hypotheses. In Comparative Politics, 26:4, 253-280.
Berg-Schlosser, Dirk (2002): Macro-quantitative vs. macro-qualitative methods in the social
sciences testing empirical theories of democracy. Paper presented at the International
Sociological Association World Congress, Brisbane.
Brayton, Robert K. And Sunil P. Khatri (1999): Multi-value Logic Synthesis. Paper. University
of Berkeley, CA. Paper Online in Internet: http://www-cad.eecs.berkeley.edu/Respep/
Research/mvsis/doc/mvsis_main.pdf [cited 2003-08-22].
Cronqvist, Lasse (2002): How MVQCA works. A short Introduction to the Ideas of the
Algorithm used in


 

. Powerpoint-Presentation. Online in Internet:

http://staff-www.uni-marburg.de/~cronqvis/tosmana/resources/ [cited 2003-08-22].

Cronqvist, Lasse (2003a): Tosmana Tool for small-n analysis. Version 1.0. Marburg:
University of Marburg. Online in Internet: http://www.tosmana.net/ [cited 2003-08-22].
Cronqvist, Lasse (2003b): Tosmana User Manual. Marburg: University of Marburg. Online in
Internet: http://www.tosmana.net/ [cited 2003-08-22].
De Meur, Gisle and Benoit Rihoux (2002): L'
analyse quali-quantitive compare (AQQCQCA): approche, techniques et applications en sciences humaines. Louvain-La-Neuve.
Drass, Kriss A. and Charles C. Ragin (1992): Qualitative Comparative Analysis 3.0. Evanston,
Illinois: Institute for Policy Research, Northwestern University. Online in Internet:
http://www.fsqca.com [cited 2003-08-22].
Goldthorpe, John (1991): "Current Issues in Comparative Macrosociology." In Comparative
Social Research, 16, 1-26.
MVSIS Group: Multi-Valued Logic Synthesis. Software Release. Online in Internet:
http://www-cad.eecs.berkeley.edu/Respep/Research/mvsis/ [cited: 2003-08-22].
Pennings, Paul (2003): Beyond dichotomous explanations: Explaining constitutional control of
the executive with fuzzy-sets. In European Journal of Political Research, 42:4, 541567.

21

Lasse Cronqvist - Using Multi-Value Logic in Social Science

Ragin, Charles C. (1987): The Comparative Method. Moving Beyond Qualitative and
Quantitative Strategies. Berkeley Los Angeles London: University of California
Press.
Ragin, Charles C. (2000): Fuzzy-Set Social Science. Chicago - London: The University of
Chicago Press.
Ragin, Charles C. and Kriss A. Drass (2002): Fuzzy-Set/Qualitative Comparative Analysis
0.963. Tucson, Arizona: Department of Sociology, University of Arizona. Online in
Internet: http://www.fsqca.com [cited: 2003-08-21].
Vanhanen, Tatu (1984): The Emergence of Democracy. A Comparative Study of 119 States,
1850-1979. Commentationes Scientiarium Socialum 24. Helsinki: Finnish Society of
Sciences and Letters.

22